sivasub987/mandate-1-safr-data

Mandate-1 SAFR synthetic semantic contrasts

Fictional research data, not transaction data, customer records, or practitioner-reviewed AML ground truth. These controlled contrasts test semantic relationships in Governance-Envelope-like records: scope, instruction, trace, evidence and reviewed context.

Try the demo · Code · Model · Kaggle mirror

Contents and intended use

Each contrast pairs an exception with a similar-looking resolved/non-exception case. The input contains labelled synthetic text spans and an input digest. Targets contain an advisory control action and citation references. Seven categories cover no material semantic exception, insufficient evidence, outside mandate, influencing action, trace contradiction, instruction conflict and no reviewed analogue.

Use for prototyping semantic sensors, controlled-contrast evaluation, weak-supervision experiments and reproducible model comparisons. Never treat the target as authority to execute a financial action.

File split	Canonical rows	Typed training/evaluation rows	Role
train	720	2,520	360 original examples + 360 teacher t01 paraphrases; exception oversampling only in typed form
paraphrase_validation	360	360	Teacher t02 paraphrases; checkpoint selection, no gradient updates
dev	60	210	Template-matched development; typed exceptions oversampled
calibration	60	60	Temperature fitting after checkpoint selection
test	120	120	Exploratory frozen synthetic evaluation
ood	60	60	Authored near-OOD, not arbitrary world-level generalisation

dataset/ is the canonical format. gliner25/ is a historical directory name for the seven-way typed format used to train Laya; it does not imply that GLiNER powers the released model. manifest.json records original construction hashes; SHA256SUMS covers files actually included in this distribution. Raw teacher generation batches are excluded. The GitHub release also retains the base pre-augmentation split for baseline reproduction.

Generation and labels

The base examples were programmatically authored from fictional contrast worlds. DeepSeek Flash, thinking disabled, paraphrased train-only pairs; a separate request to the same teacher checked label self-consistency. This is not independent validation. The teacher alias is mutable; prompt hashes, API metadata and generation/self-consistency receipts document the process. No practitioner review has occurred.

The selected t01 variants receive gradient updates; t02 variants select checkpoints. Validation shares the same latent training worlds and therefore measures wording robustness, not world-disjoint generalisation. Some t02 rows retain the original split: train provenance field; use the containing filename/Hugging Face split for evaluation membership, not that inherited field.

Test and near-OOD rows were unchanged during this distillation run and not loaded for generation, gradient updates, selection or calibration. They have nevertheless been observed in earlier experiments. Do not describe them as a pristine, independently sealed final benchmark.

Loading

from datasets import load_dataset
rows = load_dataset("sivasub987/mandate-1-safr-data", split="test")
print(rows[0]["input"], rows[0]["targets"])

For exact provenance and hashes, download the JSONL files directly. Standard-library loading:

import json
from pathlib import Path
rows = [json.loads(line) for line in Path("dataset/test.jsonl").read_text().splitlines()]

Kaggle distribution: kaggle datasets download -d sivasub987/mandate-1-safr-data --unzip -p data-download. Kaggle expands the uploaded archive into data-download/data-bundle/. Run the standard-library example from that directory (or use data-download/data-bundle/dataset/test.jsonl directly). The Hugging Face layout starts at dataset/ without the data-bundle/ prefix.

Limitations and misuse

English only; narrow fictional financial operations; one primary exception per example; template/lexical leakage possible; no genuine institutional policies, transaction distributions or investigator outcomes. A citation label is synthetic annotation, not independent evidence of truth. This dataset cannot establish AML/CFT detection quality, compliance, fairness across customers, money saved or time saved.

License and attribution

CC BY 4.0. Attribute siva-sub / Mandate-1, link the repository, retain notices and indicate modifications. Third-party code/models have separate licenses; the MAS SAFR paper is referenced, not included.

Reference: MAS SAFR. Independent research; no MAS endorsement.